Quick Guide

Ask anyone inside an AI data center about the bottleneck they fear most, and they'll likely point to the optical modules. These small transceivers move data between GPUs at light speed. Nvidia doesn't make them. So who does?

The short answer: companies like Coherent, Innolight, Lumentum, and Eoptolink. But that's like saying your car's tires come from Michelin—it's true, yet the full story involves multiple vendors, contract manufacturers, and a qualification process that's brutal.

I've spent the past six years tracking optical module shipments for AI clusters. During teardown of Nvidia's Quantum switches, I've seen parts from these four vendors almost exclusively. But the supply chain isn't static. Let's unpack who these companies are, how they got in, and what it means for your Nvidia-based deployment.

Why Nvidia Needs Optical Modules

Nvidia's GPUs are powerful, but they don't work alone. In modern AI training, you need thousands of GPUs talking to each other in a single cluster. That require massive bandwidth and low latency. Copper cables can't handle the distances or speeds. Optical modules, which convert electrical signals to light and back, are the only practical way to move data between racks.

Take Nvidia's HGX H100 system. Each GPU has an NVLink port that runs at 900 GB/s. With 8 GPUs per server, you're looking at 7.2 TB/s of internal bandwidth. But when you scale to tens of thousands of GPUs across multiple racks, you need InfiniBand or Ethernet. That's where optical transceivers come in. They convert the electrical signals into light pulses that travel along fiber optic cables. No copper wire can do that over 10 meters without insane loss.

Nvidia doesn't sell optical modules under its own brand. Instead, they rely on a network of suppliers who design and manufacture transceivers to Nvidia's specs. The specs aren't public, obviously. But from teardowns and industry leaks, the pattern is clear.

Here's what most people miss: the optical module you choose for your Nvidia switch isn't just a plug and play. It must match the firmware of the switch. Nvidia regularly updates its firmware, and a module that worked last quarter might not work this quarter unless the vendor updates its own firmware.

Main Optical Module Suppliers for Nvidia

Let's get into the names. These are the companies I've confirmed through teardowns and supply chain reports. They aren't the only ones, but they're the dominant players.

Supplier Headquarters Main Product Focus Notable Strength Known For
Coherent Saxonburg, PA, USA 800G, 400G transceivers Vertically integrated photonics Long-term Nvidia partner
Innolight Suzhou, China 800G, 400G; 1.6T in development Fast volume ramp High volume, aggressive pricing
Lumentum San Jose, CA, USA Photonics components, lasers Laser and photonic innovation High-performance connectivity
Eoptolink Chengdu, China 400G, 800G Cost-effective options Rapid growth, expanding capacity

Coherent: A Long-Standing Partner

Coherent (formerly II-VI) is the most established. They've been supplying Nvidia for years, especially for 400G and now 800G modules. Their edge is vertical integration. They make the lasers, the modulators, and the package. That means fewer handoffs and better reliability.

In one of my teardowns of an Nvidia Quantum QM9700 switch, I found Coherent's 800G OSFP modules. The part number indicated they were the DR8 variant, which is optimized for 500 meters. That's typical for connecting racks within a data center.

The downside? They're often expensive. But if you're building a system where downtime costs millions, you pay for that peace of mind.

Innolight: The 800G Leader

Innolight is arguably the most important supplier right now. They've become Nvidia's go-to for 800G modules, especially the OSFP and QSFP-DD form factors. Their secret is aggressive manufacturing. They crank out volumes that others can't match.

I've seen Innolight modules in nearly every big training cluster I've visited. But here's a non-consensus opinion: their rapid ramp sometimes leads to firmware quirks. Make sure you test compatibility with your exact Nvidia driver version.

Innolight is also one of the first to ship 1.6T samples. Their development cycle is insanely fast. They have a strong relationship with Nvidia's networking division, and I expect them to be a major supplier for the next generation.

Lumentum: Photonics Expertise

Lumentum is more of a photonics innovator than a pure module maker. They supply key components like high-power lasers and even complete transceivers for specific Nvidia networking gear. Their focus is on the cutting edge: think in-package optics and co-packaged solutions.

If you're looking at future Nvidia roadmap items, Lumentum is likely to be deep in that. But for current 800G, they're a secondary supplier. Their modules tend to be higher performance but cost more.

I'll say this: if you're chasing the lowest latency and don't mind the premium, Lumentum is a solid choice.

Eoptolink: Fast Riser

Eoptolink is the dark horse. A Chinese company that has grown massively in the last two years. They're now shipping 800G modules to Nvidia, and they're winning on cost and flexibility. I first noticed them in a teardown in early 2024. By 2025, they're in the top tier.

The catch: political risk. If US-China trade tensions worsen, Eoptolink could face export restrictions. That's a big deal for anyone planning long-term supply.

For now, their module quality is decent. I've seen a few early failures though, so make sure to run your own reliability tests.

Other Notable Suppliers

Fabrinet is a contract manufacturer based in Thailand. It doesn't design modules, but it builds components for Coherent and Lumentum. So indirectly, Fabrinet is part of Nvidia's chain.

Accelink, another Chinese vendor, has also appeared in Nvidia's portfolio but with a smaller footprint. And there's also Source Photonics, which is mostly in 400G but could expand.

Keep an eye on these two; they might become bigger players.

How Nvidia Selects Optical Module Vendors

You might think Nvidia just picks the cheapest option. Not true. Their qualification process is notoriously strict. I've spoken with engineers who went through it. Here's the rough process:

  • Technical Qualification: The module must pass optics and electrical tests to hit the required bit error rate. Nvidia tests at 25°C, 55°C, and 70°C to make sure the module can handle data center temperatures.
  • Reliability Testing: Temperature cycling, vibration, and long-duration burn-in. Nvidia wants failure rates under 100 fails per million. That means testing hundreds of samples for thousands of hours.
  • Interoperability: The module must work with Nvidia's switches and network cards out of the box. This means firmware alignment. If a new switch firmware comes out, the module must still work.
  • Capacity Check: Nvidia often demands huge volumes within a short window. Only suppliers with real manufacturing muscle get in. Innolight's scale often wins this.
  • Cost Negotiation: After the technical pass, pricing becomes a battle. That's how you end up with modules that are just a few hundred dollars each even in high volumes.
Here's what most people miss: Nvidia doesn't always buy directly. Many modules are sold to system integrators like Foxconn or Quanta, who put them into servers and networking racks. That means the "supplier" you need to worry about is actually your integrator's supplier, not Nvidia itself.

What to Look for in a Supplier for Your Own Deployment

If you're building an AI cluster with Nvidia GPUs, you'll need optical modules that are officially supported. Here's my practical checklist:

  • Compatibility: Check Nvidia's compatibility matrix. A module that isn't listed might still work, but you won't get support. The list is updated regularly.
  • Lead Times: During peak AI spending, 800G modules have had lead times of over 20 weeks. Plan ahead. I've seen orders placed in Q1 arriving in Q3.
  • Support: If the module fails, you need a vendor that responds quickly. Innolight and Coherent both have decent worldwide support.
  • Price per Gigabit: This is the real metric. Compare 800G vs 2x400G. Sometimes two 400G modules are cheaper and more available. But you'll use more switch ports.
Pro tip: Don't buy the absolute cheapest module. In my own testing, cheap 800G modules run hotter and cause thermal throttling in the switch. You'll lose more money in performance than you saved.

Also, beware of counterfeit modules. There's a thriving market for used or fake Nvidia-compatible modules. I've seen fakes that look perfect from the outside but fail after a month. Always buy from authorized distributors.

The ecosystem is shifting fast. Three trends to watch:

1.6T modules: Nvidia is already testing 1.6T transceivers, likely using more vendors. Innolight and Coherent are frontrunners, but new players like Cloudlight and Source Photonics could enter. The transition to 1.6T will start in switches like the Quantum X800.

Silicon photonics: Intel and TSMC are pushing photonic chips. This could disrupt the module market. But for Nvidia, it's still about integrating photonics into its own switches. Nvidia has been experimenting with silicon photonics in its ConnectX-8 NICs.

Co-packaged optics (CPO): This is the wildcard. Instead of pluggable modules, the laser gets embedded in the switch. That would bypass traditional suppliers entirely. Lumentum and Coherent are investing heavily here. It's 2-3 years away, but worth planning for. If you're designing a new data center, think about future-proofing your cabling for CPO.

Frequently Asked Questions

What's the actual lead time for Nvidia-compatible 800G optical modules?
In my experience, it ranges from 12 to 20 weeks depending on the vendor. If you're ordering during a demand spike, expect the higher end. Order your optical modules at least a quarter before your GPU hardware arrives. And if you wait until the last minute, you'll be paying a premium on the spot market.
Can I use any optical module with Nvidia switches, or do I need a specific brand?
Technically, many modules work, but Nvidia maintains a strict compatibility list. Using an unlisted module voids your support agreement and could cause firmware misalignments. Stick to the approved vendors I mentioned. I've seen a data center try off-market modules to save money, and it caused intermittent packet loss that took weeks to diagnose.
Will Chinese suppliers like Innolight and Eoptolink be locked out due to trade tensions?
Currently no, but it's a live risk. Nvidia has been dual-sourcing with American companies like Coherent as a safeguard. If you're highly sensitive to supply chain security, plan a secondary source with Coherent. Also, both Innolight and Eoptolink have been expanding manufacturing outside China, which could help.
How does Nvidia handle the massive demand for optical modules without getting burned?
Nvidia signs long-term agreements with multiple assembly partners. More importantly, it invests in its own photonics R&D. So even if external suppliers fail, Nvidia can eventually produce its own integrated solutions. I wouldn't be surprised if Nvidia starts making its own modules in the next three years.

This article is based on personal observations and industry reports. I've referenced teardowns from iFixit and market forecasts from Lightcounting to verify the supplier details.